On the market impact of wind energy forecastsOn the market impact of wind energy forecasts Tryggvi...
Transcript of On the market impact of wind energy forecastsOn the market impact of wind energy forecasts Tryggvi...
General rights Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.
Users may download and print one copy of any publication from the public portal for the purpose of private study or research.
You may not further distribute the material or use it for any profit-making activity or commercial gain
You may freely distribute the URL identifying the publication in the public portal If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately and investigate your claim.
Downloaded from orbit.dtu.dk on: Nov 13, 2020
On the market impact of wind energy forecasts
Jónsson, Tryggvi; Pinson, Pierre; Madsen, Henrik
Published in:Energy Economics
Link to article, DOI:10.1016/j.eneco.2009.10.018
Publication date:2010
Document VersionEarly version, also known as pre-print
Link back to DTU Orbit
Citation (APA):Jónsson, T., Pinson, P., & Madsen, H. (2010). On the market impact of wind energy forecasts. EnergyEconomics, 32(2), 313-320. https://doi.org/10.1016/j.eneco.2009.10.018
On the market impact of wind energy forecasts
Tryggvi Jonsson∗,a,b, Pierre Pinsona, Henrik Madsena
aDTU Informatics, Technical University of Denmark, Richard Petersens Plads,
building 321, DK-2800 Kgs. Lyngby, DenmarkbENFOR A/S, Lyngsø Alle 3, DK-2970 Hørsholm, Denmark
Abstract
This paper presents an analysis of how day-ahead electricity spot prices
are affected by day-ahead wind power forecasts. Demonstration of this
relationship is given as a test case for the Western Danish price area of
the Nord Pool’s Elspot market. Impact on the average price behaviour
is investigated as well as that on the distributional properties of the
price. By using a non-parametric regression model to assess the effects
of wind power forecasts on the average behaviour, the non-linearities
and time variations in the relationship are captured well and the ef-
fects are shown to be quite substantial. Furthermore, by evaluating the
distributional properties of the spot prices under different scenarios, the
impact of the wind power forecasts on the price distribution is proved to
be considerable. The conditional price distribution is moreover shown
to be non-Gaussian. This implies that forecasting models for electric-
ity spot prices for which parameters are estimated by a least squares
techniques will not have Gaussian residuals. Hence the widespread
assumption of Gaussian residuals from electricity spot price models is
shown to be inadequate for these model types. The revealed effects are
likely to be observable and qualitatively similar in other day-ahead elec-
tricity markets significantly penetrated by wind power.
Key words: Electricity market, Spot prices, Wind power forecasts,
Locally weighted polynomial regression, Conditional distributions
∗Corresponding author
Email addresses: [email protected]/[email protected] (Tryggvi Jonsson),
[email protected] (Pierre Pinson), [email protected] (Henrik Madsen)
Preprint submitted to Working Paper October 23, 2009
1. Introduction
Since the beginning of the nineties, electricity markets around the
world have undergone drastic reforms, resulting in a more deregulated
structure. The backbone of these changes has been the adoption of
wholesale electricity markets, in which producers and distributors bid
for purchase and sale of electricity. Commonly, these bids are placed
through a central clearing mechanism which determines a spot price
at which electricity is traded. However due to the complex nature of
the commodity in question, the dynamics of these spot prices are only
partially understood, making them difficult to accurately forecast. Nev-
ertheless, understanding of the price dynamics (and the resulting in-
creased predictability) is paramount for all market participants and
regulators, for the purpose of planning, trading, risk management or
alternatively market design (see e.g. Daneshi and Daneshi, 2008).
The complex nature of spot prices arises from numerous causes.
First of all, anti-gaming policies along with the instantaneous nature
of electricity and restrictions on its transmission make arbitrage over
time and space nearly impossible (Boogert and Dupont, 2005; Sewalt
and de Jong, 2003). Secondly, demand for electricity, seen from a short-
term perspective, is highly inelastic and has distinctive and complex
characteristics in the first two moments, (see e.g. Taylor and McSharry,
2007 or Panagiotelis and Smith, 2008 and references therein). Thirdly,
the electricity supply function is discontinuous, convex and steeply in-
creasing at the high demand end (Nord Pool Spot AS, 2006a; Karakat-
sani and Bunn, 2008). In parallel, the presence of non-dispatchable
renewable energy sources causes frequent variations in the shape of
the supply function, due to their low marginal costs and potential pri-
oritisation (Giabardo and Zugno, 2008). In addition, market design is
generally complex and frequently changing support schemes for some
plant technologies often lead to modifications of the market design as
well. Finally, the oligopolistic structure in many markets has given rise
to a debate about to what extent market power is exercised. Although
controversial, evidence of market power being put to force has been doc-
umented on many major electricity markets (see e.g. Eggertsson, 2003;
Schwarz et al., 2007; Christensen et al., 2007 and also Karakatsani and
Bunn, 2008 for other sources). A main objective of the present paper is
to demonstrate that wind power forecasts have an impact on the mar-
ket, and to describe how they quantitatively affect prices on the elec-
tricity market.
2
The combination of all the factors listed above results in a price be-
haviour unlike what is observed for most other traded commodities as
the price time-series often exhibits periodicity, inter- and intra-day cor-
relations, trends, mean reverting spikes, positive skewness and heavy
tails (see e.g. Conejo et al., 2005; Panagiotelis and Smith, 2008; Kosater
and Mosler, 2006, for empirical evidence of this). Furthermore, the in-
creased emphasis on renewable energy sources around the world has
made the dynamics of spot prices even more complex. The price char-
acteristics have become more extreme and made prices even harder to
predict - partly due to the very volatile nature of many of these energy
sources.
Several papers have been dedicated to describing the short-term dy-
namics of electricity spot prices, either by (i) relying solely on previ-
ous values, i.e. in a univariate time-series modelling framework (Huis-
man et al., 2006; Conejo et al., 2005; Cuaresma et al., 2004), (ii) by
accounting for the price’s response to demand, fuel prices or weather
conditions (Vehvilainen and Pyykkonen, 2005; Ruibal and Mazumdar,
2008; Mandal et al., 2006; Nogales and Conejo, 2006), or (iii) by us-
ing regime-switching approaches (Kosater and Mosler, 2006; Gonzalez
et al., 2005). However as correctly stated by Karakatsani and Bunn
(2008), the models presented in those papers have a number of limita-
tions. Firstly, fuel prices and weather conditions only affect the supply
function indirectly and their influence on the elements of the supply
function is highly non-linear. Therefore, those factors ought to be sup-
plemented by, or transformed into, information that more directly af-
fects the supply function and the behaviour of market participants in
general, as suggested by Karakatsani and Bunn (2008) and Longstaff
and Wang (2004). Secondly, most research works have been focused
on daily averages or baseload/peakload averages which conceal to some
or full extent the distinct intra-day variations of the prices. Recently,
higher frequency analysis have appeared though (e.g. Huisman et al.,
2006; Longstaff and Wang, 2004; Karakatsani and Bunn, 2008).
Operationally, energy produced by non-dispatchable energy sources
is commonly bid into the markets using forecasts of the future produc-
tion. In the case of wind power, production forecasting is a major and
rapidly growing research field and has been the topic of countless pa-
pers (see e.g Giebel et al., 2003; Costa et al., 2008, for a state of the
art review). Some efforts have been made to optimise the bidding of
wind energy into deregulated electricity markets based on these fore-
casts, and in some cases on information about their situation-dependent
3
uncertainty as well (Pinson et al., 2007; Matevosyan and Soder, 2006;
Bathurst et al., 2002). These studies however have all regarded wind
power as a price taker, and therefore not considered its potential effects
on prices. Although this approach is suitable when concentrating on in-
dividual wind power generators, wind power as a whole is in fact a price
maker as originally suggested by Skytte (1999) and Morthorst (2003),
due to its extremely low marginal cost. In areas where wind power has a
significant share in the generation portfolio, the most substantial short-
term changes in the global supply function arise from variations in wind
power generation. As shown by Giabardo et al. (2009), estimated future
wind power generation1 appears as a stochastic threshold in the sup-
ply function. The present paper presents an analysis of how electric-
ity spot prices, for the case of the Western Danish price area (DK-1) of
Nord Pool’s Elspot market, are affected by wind power forecasts. The
analysis puts emphasis on the effects of such forecasts on the mean be-
haviour of the prices, on the intra-day variations of these effects, as well
as on the corresponding impact on the distributional characteristics of
day-ahead electricity prices. Some studies have been presented on the
effects of actual power generated by wind turbines on the spot prices
in the area (Skytte, 1999; Morthorst, 2003; Moesgaard and Morthorst,
2008; Enevoldson et al., 2006) and they have shown this effect to exist
— as also expected by economical arguments. However, these studies
have been bounded to linear effects on the mean behaviour and have
therefore neither captured the full extent of this impact nor any of the
distributional effects. Furthermore, as the analysis are carried out on
the actual measured power output they only show the presence of a
relation between wind power generation and spot prices, rather than
proving wind power to be a price maker on the market. Consequently,
the resulting models can not be used for forecasting.
Showing wind power as a price maker in a short-term perspective,
implies raising the question: “What will the electricity spot prices be, if it
is believed that the wind will/will not blow?” As an answer to this ques-
tion, the relationship between wind power forecasts and spot prices is
shown not only to exist, but also to be highly non-linear and time de-
pendent. Furthermore, it is demonstrated that the correlation between
demand and wind power forecasts should not be neglected since it is
1And thereby wind power capacity due to the low marginal costs and sometimes
prioritisation of wind power
4
in fact the ratio between the forecasted wind power generation and the
forecasted load that has the strongest association with the spot prices.
In parallel, this proportional contribution of wind power to the supply
is shown to have substantial influence on the distribution of the spot
prices as well.
For the demonstration of these effects, a non-parametric regression
model is employed. The relationship between forecasted wind power
production and electricity spot prices is estimated for every hour of the
day. This dependency is estimated by assuming that the relationship
can be locally described with a second order polynomial. For estima-
tion, a least squares criteria is employed allowing for the mean effect
of forecasted wind power penetration on the prices to be extracted. De-
spite the facts listed about the (non-Gaussian) properties of the condi-
tional spot price distribution, this approach serves well for estimating
the mean behaviour of the spot prices, partially due to the non-linear
properties of the model and the large amount of data used as input to
analysis. Employing such a criteria for the non-parametric regression
model prevents however from any inference on the effects of wind power
forecasts on distributional properties of spot prices to be made. There-
fore, the evaluation of these effects is performed by directly analysing
the first four moments of the price distributions under different scenar-
ios of the ratio between forecasted wind power and forecasted load. The
wind power forecasts used are generated by WPPT2 (see Nielsen et al.,
2002), a wind power forecasting tool that has been successfully used in
Denmark over the past years and was used for bidding aid by almost
every wind turbine owner in DK-1 during the period in question.
The remainder of the paper is structured as follows: Section 2 de-
scribes the market structure at Elspot and the data set used as input to
the analysis. Section 3 provides a brief introduction to the mathemati-
cal approach to the modelling. In Section 4, an analysis of how the mean
behaviour of the spot prices is affected by wind power forecasts is given,
while the impact of these forecasts on the price distribution is the topic
of Section 5. Finally, concluding remarks are given in Section 6 along
with some general discussion.
2The Wind Power Prediction Tool
5
2. Nord Pool’s Elspot
2.1. Market Setting
The analysis presented in this paper is done on data from the West-
ern Danish price area (DK-1) of Nord Pool’s Elspot market. Elspot is a
day-ahead physical delivery market for electricity. It is currently oper-
ating in the entire Scandinavia (Denmark, Finland, Norway and Swe-
den) and in the so called Kontek area, located in Northeastern Germany.
At Elspot, spot prices are set by a market equilibrium model, where
supply and demand curves of all market participants are matched on a
day-ahead basis. Gate closure is at noon each day for the period mid-
night to midnight in the following day and prices are published later
in the day with a resolution of one hour. The one hour prices are cal-
culated by matching the collaborative supply and demand curves, cal-
culated from the bids and ask prices placed by the market participants
(see Nord Pool Spot AS, 2006a,b, for details). This price, found from bids
from all market participants, defines the system price from which the
area prices are defined.
Due to transmission constraints, the region covered by Elspot is di-
vided into several price areas. Physical constraints on transmission de-
fine the outer bounds of each price area, implying that transmission
capacity within an area can be regarded as unlimited. The area spot
prices are calculated in the same manner as the system price, but only
considering the bids within the area along with possible utilisation of
the transmission lines to surrounding areas (see Nord Pool Spot AS,
2006b, for details). The area prices, which determine at what price
physical trading is done within an area, can therefore differ quite con-
siderably between areas. If none of the interconnections between areas
are fully utilised, the system price is valid in the whole region. However,
this is seldom the case and therefore is modelling of the area spot prices
appropriate when short term dynamics and forecasting are considered.
The DK-1 price area consists of Jutland, Funen and the islands west
of the Great Belt. The area is an interesting context for investigation
of electricity spot prices as it can be said to represent the future of
liberalised electricity markets. This is because it has relatively large
connections to its surrounding areas, and is heavily penetrated by an
inexpensive, non-dispatchable energy source, i.e. wind power. In fact,
DK-1 is currently the grid area in the world that has the largest share of
wind power in its generation portfolio, with more than 20% of its annual
consumption generated by wind turbines.
6
2.2. The Data
The data, which the analysis here presented is carried out on, covers
the period from January 4th 2006 to October 31st 2007. It consists of
hourly area spot prices along with hourly consumption measurements
for the area3 and wind power forecasts (in MW) with a temporal reso-
lution of 15 minutes, made at 07:00 on the day before delivery for lead
times up to 48 hours. The forecasts are made using WPPT (see Nielsen
et al., 2002).
If the analysis is to be true to the criteria of analysing future elec-
tricity prices, both measures entering the wind power-load ratio have
to be forecasts instead of actual measurements. When load forecasts
are made using state of the art load forecasting models (e.g. Taylor and
McSharry, 2007), the relationship between the actual load and the pre-
dicted load can be described as
Lt = Lt + εt where εt ∼ N(0,σ2) (1)
where Lt is the actual load, Lt is the predicted load and σ2 is the fi-
nite variance of the residuals, εt. Hence, by adding a Gaussian noise
with the appropriate variance to the load measurements, a time series
that has the characteristics of an actual load forecast series is obtained.
The standard deviation of the noise is chosen as 2% of the average load
for the period, since it reflects the performance of state of the art load
forecasting models (see e.g. Taylor and McSharry, 2007). The use of
simulated load forecasts gives rise to some deviations from the real-
life situation though. First of all, the residuals of actual load forecasts
are bound to have some autocorrelation in the lags up to the prediction
horizon. This is not reflected in the simulated residuals. However, for
forecasts of such a degree of accuracy as load forecasts are in general,
the small prediction error is reflected by a small residual autocorrela-
tion as well. The influence of the missing autocorrelation structure is
therefore only marginal. Secondly, since both load forecasts and wind
power predictions are typically based on weather forecasts, some cor-
relation between the errors of the two might be observed in practise.
This is however not the case when simulated forecasts are used since
the computer generated noise, added to the load forecasts, is not corre-
lated to the real error of the wind power predictions. The error of the
3Available at http://www.energinet.dk/en/menu/Market/Market.htm and
at http://www.nordpoolspot.com
7
wind-load forecasting ratio will therefore have characteristics that dif-
fer slightly from what would be observed if both were actual forecasts.
Nevertheless, the mean of the errors will still be zero in both cases and
other effects are minor. These potential dissimilarities from the prac-
tical situation are ignored in the analysis to follow due to their small
impact on the results.
Hourly wind energy forecasts in MWh also have to be derived. This
is done by linearly interpolating between each two adjacent forecasts in
every hour and taking the result as the production in MWh for that 15
minute period. These interpolations are then summed up for each hour.
So in mathematical terms, an hourly forecast is obtained by
V(h)t =
5
∑i=2
0.25 ·
V
(q)t,q(i−1)
+ V(q)t,q(i)
2
(2)
where V(h)t (hereafter noted as Vt) is the hourly wind energy forecast
for hour t, and V(q)t,q(i)
is the 15 minute wind power forecast for quarter i
within hour t.In order to account for the correlation between demand and wind
power forecasts, the wind power penetration level, V(p)t , is defined as
V(p)t =
Vt
Lt
, (3)
Finally, it should be emphasised that no extreme events are excluded
from the data set.
3. Non-parametric regression modelling of day-ahead electric-
ity prices
The relationship between, area spot price and wind power genera-
tion is far from being linear. It is therefore essential for obtaining a
proper estimate of the effects of wind power forecasts on the area spot
price to account for these non-linearities. The problem of estimating a
complex non-linear relationship between variables is however not a triv-
ial one. But by assuming that the relationship is locally linear or locally
describable by a low order polynomial the problem is much more con-
venient to deal with. One could locally solve a weighted least squares
problem as described in detail for the general case in (Cleveland and
Devlin, 1988; Nielsen et al., 2000; Madsen and Holst, 2000).
8
Let a model for the spot prices at time t, Pt, be defined as
Pt = θ(xt) + εt, (4)
where θ(·) is a vector of coefficient functions and εt is a noise term.
Furthermore, xt is a vector of explanatory variables. In this case those
variables are some direct or derived form of a wind power forecast, Vt,
and an hour of the day indicator, kt.
The functions θ(·) are estimated at a number of distinct points by
approximating the functions using polynomials and fitting the resulting
linear model locally to each of these fitting points. More specifically, let
U =[
VU kU
]Tdenote a particular fitting point, chosen from a set of
m total fitting points, and let p2(U) be a column vector of terms in the
corresponding 2nd-order polynomial, i.e.
p2(U) =[
1 VU kU V2U VUkU k2
U
]T. (5)
Furthermore, let φU =[
φU,1 . . . φU,7
]Tdenote the coefficient vector
at U. Now the linear model
PU = p2(U)TφU + εU (6)
can be used to describe the spot price in the close vicinity of U and can
be fitted locally using weighted least squares (WLS), i.e.
φ(U) = argminφU
N
∑i=1
wU(xi)(
Pi − xTi φU
)2(7)
where xi = p2(Ui) and for which a unique closed-form solution exists
provided that the matrix with rows xi corresponding to non-zero weights
has a full rank. The weights are assigned as
wU(xi) = W
(||xi − x||2
h(x)
)(8)
where W(·) is a decreasing weight function taking non-negative argu-
ments. Furthermore, h(x) is the bandwidth used for the particular fit-
ting point, i.e. the maximum euclidean distance between a fitting point
and an observation resulting in a non-zero weight being assigned in
Eq. (7). This implies that a small bandwidth will result in a very flexi-
ble model with low bias and high variance while applying a large band-
width yields a more rigid model having higher bias but lower variance.
9
For readers familiar to exponential smoothing, applying a small band-
width corresponds to having a low level of smoothing in the model. The
bandwidth can therefore be said to be a scalar controlling the rate at
which the weight of an observation, in the estimation, decreases with
distance from the fitting point. From this it follows that the argument
which W(·) takes is the relative distance between the fitting point and
the other point falling within the bandwidth. Following Cleveland and
Devlin (1988) and Nielsen et al. (2000), a tri-cube kernel is chosen as a
weight function so
W(u) =
{(1 − u3)3 u ∈ [0,1)
0 u ∈ [1,∞). (9)
Turning back to the global view on the model, it can be seen that for
an arbitrary chosen U, out of a set of m fitting points, there can be found
a parameter vector φU. From these, the elements of θ(xt) are estimated
as
θ(xt) = θ(p2(U = xt)) = pT2 (U)φ(U) (10)
where φ(U) is the WLS estimate of φU .
The bandwidth, h(x) is chosen so that at any given time 30% of all
observations fulfil ||xi − x||2 ≤ h(x). In other words, the bandwidth is
varied according to the local density of the data by letting h(x) be equal
to the distance of the qth-nearest xi to x, where q is 30% of the total
number of observations (see e.g. Cleveland and Devlin, 1988, for de-
tails). The choice of the criteria that 30% of the observations should fall
within the bandwidth is made since it was desired to obtain as local es-
timates as possible and 30% was the smallest bandwidth that resulted
in a full rank design matrix at all times. This criteria is applied for
estimation in all m = 242 fitting points.
Despite what has been stated previously in this paper regarding
the conditional distribution of the spot prices, using this sort of least
squares technique is deemed suitable for this analysis. This is because
the model is only used for assessing the average dependency between
the variables and for such estimates, Gaussian estimators are gener-
ally known to be the best ones for most types of data. Furthermore, the
non-Gaussianity of the residuals is reduced by the model’s non-linearity.
This aside, as will be demonstrated later on in the paper (Figure 5), de-
spite the prices not being normally distributed, their distribution has
a bell shaped form, making the Gaussian assumption not completely
10
inappropriate. In addition, the analysis is carried out on a very exten-
sive data set containing hourly observations for 22 consecutive months.
Therefore, will the impact of each individual extreme only be minimal.
Hence, the use of robust least squares or similar techniques is not nec-
essary.
How the method is applied in the analysis to follow can be sum-
marised as follows. First the data is scaled so all variables are between
[−1,1]. Then a grid of m = 24 × 24 equidistant fitting points is defined.
For each of these fitting points, Umi, the following steps are taken:
1. Calculate the euclidean distances between every observation and
the fitting point of interest. Applying the bandwidth principle pre-
viously described, these distances are normalised by that corre-
sponding to the qth-nearest neighbour, where q is set to 30%, and
thereby form the input to Eq. (9).
2. Compute the estimate Φ(Umi) of the local polynomial coefficients
at the fitting point Umiby solving Eq. (7).
3. Obtain the local estimate of the spot price at Umifrom Eq. (10)
with p2(Umi) as the polynomial for the particular fitting point as
described by Eq. (5).
The mean spot price for any point U can be obtained by bilinear inter-
polation from the local estimates calculated at each fitting point. This
finally yields a smooth trend surface like those shown and commented
on in the following.
4. General trend: the effect of forecasts on the mean price
On a day-ahead basis, the area spot price is subject to a considerable
uncertainty. It can therefore rightfully be stated that the future spot
price has some unknown distribution and the model presented in the
previous section can be used, if applied correctly, to provide information
about the mean in this distribution.
In Figure 1 the average spot price in DK-1 is estimated as a function
of both the time of the day and the forecasted wind energy production
measured in MWh per hour. From the figure, it is quite obvious that
forecasts of large wind power production in a given hour will, on aver-
age, result in a lower spot price in that hour, since when going along the
wind power-axis, in the increasing direction, the mean price decreases.
11
05
1015
20
0
500
1000
1500
20
40
60
HourForecasted WP production
Pric
e [E
UR
/MW
h]
20
25
30
35
40
45
50
55
Figure 1: The dependence of the spot prices on forecasted wind power production and
its variation throughout the day
During the night, the average price varies from around 30AC/MWh for a
forecasted low wind power production down to around 18AC/MWh for
the hours with forecasted large wind power production. During the
day, this difference between the two extremes in forecasted quantities of
wind power produced is somewhat larger as the prices go from around
50 − 55AC/MWh down to around 30AC/MWh. Due to the infrequent oc-
currence of the installed wind power capacity being fully utilised, the
wind power-axis only reaches 1500 MWh which is somewhat lower than
the full utilisation4. Observations above the 1500 MWh limit are never-
theless used for estimation when it is relevant.
What is also very interesting is that the daily price raise, during
the hours of the day where consumption reaches its daily peak, evens
out as wind power production in the system increases. The reason for
this is that the virtually nil marginal cost of the wind turbines shifts
the supply curve to the right when more wind power is produced and
therefore it takes more consumption to reach the steep end of it. In other
4Installed wind generation capacity in DK-1 was around 2400 MW during the con-
sidered period.
12
words, the increased production of the turbines means that less cost
efficient plants otherwise covering the base load, will be covering the
peaks only. This in turn prevents the even less cost efficient generators
from being utilised during the hours when demand peaks.
As stated earlier, the demand for electricity varies severely through-
out the day and the week. The same quantity of wind power, measured
in MWh, can therefore have quite different effects on the price depend-
ing on the time of the day and the week. More specifically, electricity
demand is generally lower during the evening and the night and there-
fore will a large volume of wind power produced in these hours make up
for a larger share of the total demand than the same quantity would do
during the day. This in turn will cause the equilibrium point of the sup-
ply and demand curves to be placed lower during the evening and the
night than it would during the day. In order to eliminate these effects,
to some extent at least, the wind power forecasts can be included as
the proportional contribution to the total supply instead of its absolute
contribution. In other words, by substituting the forecasted wind power
production, Vt, with the forecasted wind power penetration defined in
Eq. (3), V(p)t , a better prediction of the price equilibrium is gained from
the wind power forecasts.
There are certainly other ways of deriving a number representing
the interaction between wind power predictions and load forecasts. For
instance, the difference between the forecasts could be considered in-
stead of the ratio between the two. Although the relationship will then
appear differently, simulations indicate that the extent of the impact
will be approximately the same. It is also intuitively appealing to work
with values that are between 0 and 1. Therefore, further analysis is
only presented on V(p)t as it is defined in Eq. (3).
In Figure 2 a smooth estimate of the spot price, as a function of
the time of the day and forecasted wind power penetration, is given.
The figure shows the same type of effects as described before. How-
ever, the effects are more dramatic than seen in Figure 1 since the same
actual production now has different effects depending on what time of
the day and the week5 the production occurs. The average spot price
is again considerably lower at times where wind power production has
been predicted to be large. The difference between the two extremes
in forecasted production is roughly the same during night hours, while
5Due to the weekly variation in the load
13
05
1015
20
00.2
0.40.6
20
30
40
50
60
HourForecasted WP penetration
Pric
e [E
UR
/MW
h]
20
25
30
35
40
45
50
55
60
Figure 2: The dependence of the spot prices on forecasted wind power penetration and
its variation throughout the day
the difference has increased during the day. During the day, the average
spot price is around 55− 60AC/MWh when nothing of the demand is sup-
plied by wind power. The average prices then rapidly diminishes with a
small increase in wind power penetration, and after a short stand still
as the penetration approaches 20%, the sharp decline continues up to
around 40% predicted penetration, for which the average spot price is
around 35AC/MWh. When the forecasted wind penetration has reached
40%, a decrease in average price per penetration percent becomes more
subtle and as the forecasted wind power penetration reaches 80%, the
average spot price has declined to around 22 − 25AC/MWh.
Some general information about the characteristics of the supply
function can also be deduced from the figure. The rather sharp gradient
changes in the average price in the lower penetration end of the plot
are a clear indication of the discontinuity of the supply function. Wind
power quickly pushes the steepest end of the supply curve to the right
side of the equilibrium point, and thereby out of the generation port-
folio for that hour, explaining the sharp decline in the average price.
After stabilising itself, the average price decreases rapidly again when
another threshold is pushed out of the equilibrium. The final subtle de-
cline, and the night time behaviour can also be explained by considering
14
<10 11−20 21−30 31−40 41−50 51−60 61−70 >700
10
20
30
40
50P
rice
[EU
R/M
Wh]
Wind penetration [%]
Figure 3: Average spot price, categorised by intervals of forecasted wind power pene-
tration, in DK-1 in the period January 4th 2006 - October 31st 2007
the shape of the underlying supply function, since the equilibrium point
has then reached the flatter part of the supply function where cheaper
energy sources such as Norwegian hydro power and wind power are
placed.
Even though it is clear from Figure 2 that forecasted wind power
does in fact influence the spot prices, it is hard to point out how big the
actual effects are. In Figures 3 and 4 the extent of the effects are better
illustrated. In Figure 3 the average spot price is shown for predicted
wind power penetration on certain intervals in the period which the
data set spans. It shows how the average spot price generally decreases
as the share of wind power in the system increases.
In Figure 4, the impact of forecasted wind power penetration is for-
mulated in terms of reduction in price compared to no wind being present
in the system. For doing this, the assumption is made that wind power
penetration under 4%, corresponding to a production of approximately
80 MWh per hour, has very little or no effects on the spot prices. Obser-
vations falling on this interval are taken as a reference point and rep-
resents the situation when no wind power is predicted to enter the sys-
tem. Comparing the average spot price for the reference group, which
is AC44.43, to that of the remaining observations, where the average is
AC36.68, shows that the spot prices drop on average by 17.5% when the
15
4−10 11−20 21−30 31−40 41−50 51−60 61−70 >700
10
20
30
40
50
60P
rice
redu
ctio
n [%
]
Wind penetration [%]
Figure 4: Reduction in average spot price, compared to the ”no wind” situation, for
different levels of forecasted wind power penetration in DK-1 in the period January 4th
2006 - October 31st 2007
forecasted wind power penetration exceeds 4%. In Figure 4, the pene-
tration levels above 4% have been divided into intervals of 6-10% and
the bars represent how much lower on average, the price is during pe-
riods of the given penetration interval, compared to the reference ”no
wind” situation. The plot clearly illustrates that the spot prices tend to
decrease as the forecasted wind power penetration increases.
The extent and the characteristics of the effects that have been shown
to exist here indicate that properly accounting for them will be of seri-
ous help when electricity spot prices are to be forecasted in an area
penetrated by wind power to some or large extent. These effects are
consistent with what intuitively would be expected and would provide a
forecasting model with vital information about the current shape of the
supply function and thereby a good indication about the equilibrium
point. However although the existence of these effects is undisputed, it
can be debated whether they are for the good or worse, now when the
share of wind power stands to be increased all over the world. Less ex-
pensive electricity might sound appealing for many at first, especially
put in context with marginal bidding. On the other hand, this results in
less contribution to the enormous initial investments of power plants of
any kind. This will in turn reduce investors’ interest in investing in new
16
plants. Furthermore, lower electricity prices pose a threat to the exis-
tence of flexible power plants that produce electricity at high marginal
costs. These plants however contribute heavily to a much needed sta-
bility in the energy supply.
5. Effects of wind power forecasts on price distributional prop-
erties
Having established that forecasted wind power penetration certainly
affects the mean spot price in DK-1, the question remains whether it
affects the distribution of the prices as well. Equipped with more com-
prehensive knowledge about the relationship between price volatility
and one or more of its fundamental causes (in this case wind power
forecasts), one may better explain and hereby estimate future volatility
levels while conditioning it upon these causes, e.g. in a non-parametric
fashion. For carrying out the distribution analysis, the data set is di-
vided into bins, according to forecasted wind power penetration, so that
approximately 2500-3000 observations belong to each segment. Then
the properties of the price distribution are estimated within each bin.
In Figure 5, histograms of the electricity prices are shown for differ-
ent levels of forecasted wind power penetration. The figure illustrates,
what already has been established, that the mean price shifts towards
zero as forecasted wind power penetration increases. Furthermore, the
positive skewness of the price distribution is quite evident from the fig-
ure as well as the fact that the heavy tail diminishes with increased
forecasted wind power penetration. This translates to the statement
that the probability of extremely high prices is much lower when the
wind power penetration is predicted to be high.
The difference in distribution properties is summarised in Table 1.
The first two lines in the table show the shift of mean, already dis-
Table 1: Properties of the spot price distribution for different scenarios of forecasted
wind power penetration
0-5% 5-10% 10-16% 16-25% 25-40% 40-100%
Mean 42.98 41.13 40.26 38.10 33.24 26.02
Std. Dev. 16.95 15.32 14.18 13.08 11.35 11.23
Skewness 0.82 0.48 0.39 0.63 0.32 0.29
Kurtosis 4.41 3.39 3.40 4.05 3.04 4.09
17
0
5
10
Price [EUR]
Pro
port
ion
[%]
0 50 1000
5
10
0 50 100 0 50 100
0−5%5−10%10−16%16−25%25−40%40−100%
0 20 40 60 80 100 1200
2
4
6
8
10
Price [EUR]
Pro
port
ion
[%]
Figure 5: Distribution of prices for different intervals of forecasted wind power pene-
tration
18
cussed, and reduction in standard deviation, indicating less volatility
of the prices. Lines 3 and 4 show how the skewness and kurtosis of
the distributions change for the different levels of penetration. Despite
the fact that no obvious pattern is detectable in lines 3 and 4, they rep-
resent quite dissimilar distributions. Taking the penetration intervals
from left to right in the table, the first distribution is rather skewed
and with high kurtosis, due to the heavy tail seen in Figure 5. Then as
the wind power penetration increases in the 2nd and 3rd intervals, the
tail becomes not as heavy, while the mean does not shift all that much,
explaining the decrease in both skewness and kurtosis. When the pene-
tration reaches the level of the 4th interval, the mean has shifted more
while the rather high prices still occur. Hence, the increase in skewness
and kurtosis. For predicted wind power penetration between 25-40%
these extreme price situations no longer occur, reflected in a decrease
both in skewness and kurtosis. This threshold effect is yet another non-
linear effect introduced in the market by wind power or wind power
forecasts. Finally, for the highest forecasted penetration interval, the
frequency of very low prices increases substantially, explaining the re-
duction of skewness and increase in kurtosis. So to summarise, going
from a low wind power penetration to high, generally leads to a lower
skewness due to the diminishing frequency of very high prices along
with the increased probability of very low prices.
Another thing that catches the eye in Figure 5 is the relatively high
proportion of prices equal to zero in the histogram representing the
highest penetration interval. Over 2% of the times when wind pene-
tration is above 40%, electricity spot prices are 0AC/MWh. Although
ill-detectable from the figure, this situation rarely occurs for the 25 -
40% penetration interval, while it does not occur for the lower levels of
penetration. For the four lower ones, the minimum price does however
approach zero as the forecasted penetration increases. The occurrence
of the spot price being 0AC/MWh will result in a negative cash flow for
producers subject to imbalance costs in that hour. In other words, it
will pay off, even for producers with no marginal production costs, not
to produce electricity. This further supports what was stated at the end
of previous section about the down side of increased wind power pene-
tration under current market conditions.
From a modelling perspective, the dissimilarity of the spot price
distribution between different levels of forecasted wind power penetra-
tion strongly indicates that estimating prediction intervals conditioned
on the wind power predictions is worth the effort. It is quite obvious
19
that the spot prices are not Gaussian distributed and therefore it must
be deemed highly unlikely that models constructed with least squares
techniques will have Gaussian residuals. Prediction intervals for such
models should therefore be estimated using other techniques. In fact
the distributions are so far from parameterised distributions that it
seems reasonable to conclude that non-parametric approaches, like for
instance quantile regression (Møller et al., 2008), will return the most
reliable prediction intervals.
6. Conclusions and Discussion
The analysis presented in this paper demonstrates the dramatic im-
pact of predicted wind power penetration in the system on not only the
level of the spot prices but also their distributional characteristics. The
spot price is, on average, shown to decrease with increased predicted
wind power penetration, while intra-day price variations diminish to
some extent. As all this happens in a non-linear manner, the use of the
non-parametric regression model for the analysis proves to be very ben-
eficial. Furthermore, wind power forecasts are shown to cause threshold
effect in the price behaviour, with e.g. the appearance of zero prices, or
the removal of extreme prices. They could therefore contribute to an
understanding of some of the non-linearities and regime-switching be-
haviour in the prices. The results of this paper therefore support some
of the conclusions of Karakatsani and Bunn (2008) such that aspects
of plant dynamics should be considered when models of the short-term
dynamics of electricity spot prices are to be derived. It would therefore
be interesting to derive a forecasting model for these prices that ac-
counts for the impact of forecasted wind-to-load ratio (see e.g. Jonsson,
2008). When developing such an approach, accounting for the uncov-
ered non-linearities in the relationship between forecasted wind power
penetration and day-ahead electricity prices will be essential.
The findings of this paper confirm what previous studies of the im-
pact of wind power on electricity spot prices have shown, that wind
power has a non-negligible impact on day-ahead electricity prices. Here
however, based on the claim that it is instead the predicted wind power
penetration that should be seen as an explanatory variable, the impact
is shown to be more substantial than previously recorded (Enevoldson
et al., 2006; Moesgaard and Morthorst, 2008). The corresponding rela-
tionship moreover turns out to be highly non-linear, and the distribu-
tional characteristics of prices are also affected. The results also show
20
that a simple assumption like Gaussianity commonly made for the es-
timation of prediction intervals of electricity prices (e.g. Nogales and
Conejo, 2006) does not hold. The fact that the spot prices themselves
are so far from being Normally distributed will certainly be reflected in
the residual distribution of a forecasting model, which parameters are
estimated with a least squares criteria. Furthermore, as the price dis-
tribution has also been shown to be dependent on an external signal,
prediction intervals should be generated accounting for this relation,
thus yielding conditional prediction intervals. The analysis therefore
indicates that estimating the uncertainty conditioned on an explana-
tory variable, e.g. wind power forecasts, in a non-parametric fashion
could increase the resolution of probabilistic forecasts of electricity spot
prices.
In this paper, the scope has been the Western Danish price area (DK-
1) at Nord Pool. This price area may be seen as representative of the
future deregulated electricity with significant penetration of renewable
energy generation. Although the share of wind power in DK-1 is larger
than anywhere else in the world, it is very plausible that the effects,
revealed here, can also be detected in other market areas as well, pen-
etrated by wind power to some extent — for instance in Spain or Ger-
many. Similar causes would have similar effects, the principal ones
being varying availability of the fuel and extremely low marginal costs.
The severe impact of wind power forecasts on all behaviour of the
electricity prices is also interesting to consider in the context of market
design and with the long-term development of the production portfolio
in Denmark in mind — where the intention is to increase the share
of wind power generation up to 50% of the electricity consumption by
2025 (Ea Energy Analyses, 2007). With the current market structure
of marginal bidding, the frequency of hours where the spot price is
zero is bound to increase along with growing wind power penetration
in the system in a similar manner as has been demonstrated here. This
will further enhance the stochastic threshold effect demonstrated here,
and thereby increase price volatility and cause it to have alternating
weather dependent patterns (Meibom, 2007). This aside, higher risk
premium will be required on investments in all sorts of new energy
generation capacity and investment in conventional power generation
capacity will be more focused on flexibility than efficiency since those
plants will have to rely more on the increased demand on the balance
markets as a source of income (Meibom, 2007). The impact of wind
power on the price making at the electricity markets should therefore
21
be given careful thought when the possibility of increasing the share
of wind power, and non-dispatchable energy sources in general, in the
generation portfolio is discussed. Especially, the fact that wind power
penetration has some non-linear effects on the prices should be taken
into consideration, as it implies that current market situation can not
be scaled directly for analysing the future circumstances. In the con-
text of market design, it will also be interesting to monitor the market’s
response to other renewable sources reaching the status of making up
for a significant share of the energy supply. Those sources in all likeli-
hood being solar or wave energy in the medium term. Whether this will
level out the effect of increased wind power or magnify them will play
an important role in future development of the market structure.
7. Acknowledgements
The authors would like to thank Energinet.dk, the danish Trans-
mission system operator, and Nord Pool ASA for providing data for
the analysis, publicly available at http://www.energinet.dk/en/menu/Market/Market.htm and at http://www.nordpoolspot.com .
Authors are also grateful to Henrik Aa. Nielsen at ENFOR A/S for fruit-
ful discussions and useful comments and Torben Skov Nielsen at EN-
FOR A/S for providing wind power predictions.
References
Bathurst, G. N., Weatherill, J., Strabac, G., 2002. Trading wind gen-
eration in short term energy markets. IEEE Transactions on Power
Systems 17(3), 782–89.
Boogert, A., Dupont, D., 2005. On the effectiveness of the anti-gaming
policy between the day-ahead and real-time electricity markets in the
netherlands. Energy Economics 27(5), 752–770.
Christensen, B. J., Jensen, T. E., Mølgaard, R., 2007. Market power
in power markets: Evidence from forward prices of electricity. CRE-
ATES Research Paper No. 2007-30. Available at http://ssrn.com/abstract=1150145 .
Cleveland, W. S., Devlin, S. J., 1988. Locally weighted regression: An
approach to regression analysis by local fitting. Journal of the Ameri-
can Statistical Association 83(403), 596–610.
22
Conejo, A. J., Contreras, J., Espınola, R., Plazas, M. A., 2005. Forecast-
ing electricity prices for a day-ahead pool-based electric energy mar-
ket. International Journal of Forecasting 21(3), 435–462.
Costa, A., Crespo, A., Navarro, J., Lizcano, G., Madsen, H., Feitona,
E., 2008. A review on the young history of wind power short-term
prediction. Renewable & Sustainable Energy Reviews 12(6), 1725–
1744.
Cuaresma, J. C., Hlouskova, J., Kossmeier, S., Obersteiner, M., 2004.
Forecasting electricity spot-prices using linear univariate time-series
models. Applied Energy 77(1), 87–106.
Daneshi, H., Daneshi, A., April 2008. Price forecasting in deregulated
electricity markets - a bibliographical survey. In: The 3rd IEEE Inter-
national Conference on Electric Utility Deregulation, Restructuring,
and Power Technology (DRPT2008). Nanjing, China.
Ea Energy Analyses, 2007. 50% wind power in denmark in 2025 -
english summary. Available at: http://www.talentfactory.dk/media(2513,1033)/081029_50pct._wind_power_i%n_dk_in _2025.pdf .
Eggertsson, H., 2003. The scandinavian elecricity power market and
market power. Master’s thesis, Technical University of Denmark,
DTU, Kgs. Lyngby, Denmark, available at http://www2.imm.dtu.dk/pubdb/views/publication_details.php?id=2533 .
Enevoldson, S. W., Østergaard, P. A., Morthorst, P. E., Moesgaard, R.,
2006. Vindkraftens betydning for elprisen i danmark. Tech. rep., IBT-
Wind, in Danish.
Giabardo, P., Zugno, M., 2008. Competitive bidding and stability anal-
ysis in electricity markets using control theory. Master’s thesis, In-
formatics and Mathematical Modeling, Technical University of Den-
mark, DTU, Kgs. Lyngby, Denmark, available at www2.imm.dtu.dk/˜pp/thesis.htm .
Giabardo, P., Zugno, M., Pinson, P., Madsen, H., 2009. Feedback, com-
petition and stochasticity in a day ahead electricity market. Energy
Economics Available online.
23
Giebel, G., Kariniotakis, G., Brownsword, R., 2003. The state-
of-the-art in short-term prediction of wind power - a litera-
ture overview. EU project Anemos, Deliverable Report D1.1,
Available at http://anemos.cma.fr/download/ANEMOS_D1.1_StateOfTheArt_v1.1.pdf .
Gonzalez, A. M., Roque, A. M. S., Garcıa-Gonzalez, J., 2005. Modeling
and forecasting electricity prices with input/output hidden markov
models. IEEE Transactions on Power Systems 20(1), 13–24.
Huisman, R., Huurman, C., Mahieu, R., 2006. Hourly electricity prices
in day-ahead markets. Energy Economics 29(2), 240–248.
Jonsson, T., 2008. Forecasting of electricity prices accounting for wind
power predictions. Master’s thesis, Informatics and Mathematical
Modeling, Technical University of Denmark, DTU, Kgs. Lyngby, Den-
mark.
Karakatsani, N. V., Bunn, D. W., 2008. Forecasting electricity prices:
The impact of fundamentals and time-varying coefficients. Interna-
tional Journal of Forecasting 24(4), 764–785.
Kosater, P., Mosler, K., 2006. Can Markov regime-switching models
improve power-price forecasts? Evidence from German daily power
prices. Applied Energy 83(9), 943–958.
Longstaff, F. A., Wang, A. W., 2004. Electricity forward prices: A high-
frequency empirical analysis. The Journal of Finance 59(4), 1877–
1900.
Madsen, H., Holst, J., 2000. Modelling Non-Linear and Non-Stationary
Time Series (Lecture Notes). IMM, DTU, Kgs. Lyngby, Denmark.
Mandal, P., Senjyu, T., Funabashi, T., 2006. Neural networks approach
to forecast several hour ahead electricity prices and loads in deregu-
lated markets. Energy Conversion & Management 47(15-16), 2128–
2142.
Matevosyan, J., Soder, L., 2006. Minimization of imbalance cost trading
wind power on the short-term power market. IEEE Transactions on
Power Systems 21(3), 1396–1404.
Meibom, P., 2007. Market consequences [in my view]. IEEE Power and
Energy Magazine 5(6), 120–118.
24
Moesgaard, R., Morthorst, P. E., April 2008. The impact of wind power
on electricity prices in denmark. In: EWEC 2008, European Wind
Energy Conference, Business and Policy Track. Brussels, Belgium.
Møller, J. K., Nielsen, H. A., Madsen, H., 2008. Time-adaptive quantile
regression. Computational Statistics and Data Analysis 52(3), 1292–
1303.
Morthorst, P. E., 2003. Wind power and the conditions at a liberalized
power market. Wind Energy 6(3), 297–308.
Nielsen, H. A., Nielsen, T. S., Joensen, A. K., Madsen, H., Holst, J., 2000.
Tracking time-varying coefficient functions. International Journal of
Adaptive Control and Signal Processing 14(8), 813–828.
Nielsen, T. S., Nielsen, H. A., Madsen, H., 2002. Prediction of wind
power using time-varying coefficient functions. In: 15th IFAC World
Congress. Barcelona, Spain.
Nogales, F., Conejo, A. J., 2006. Electricity price forecasting through
transfer function models. Journal of Operational Research Society
57(3), 350–356.
Nord Pool Spot AS, 2006a. Bidding in Nord Pool Spot’s Elspot Market.
Available at www.nordpoolspot.com .
Nord Pool Spot AS, 2006b. Calculation of System- and Area prices.
Available at www.nordpoolspot.com .
Panagiotelis, A., Smith, M., 2008. Bayesian density forecasting of intra-
day electricity prices using multivariate skew t distributions. Inter-
national Journal of Forecasting 24(4), 710–727.
Pinson, P., Chevallier, C., Kariniotakis, G. N., 2007. Trading wind gen-
eration from short-term probabilistic forecasts of wind power. IEEE
Transactions on Power Systems 22 (3), 1148–1156.
Ruibal, C. M., Mazumdar, M., 2008. Forecasting the mean and the vari-
ance of electricity prices in deregulated markets. IEEE Transactions
on Power Systems 23 (1), 25–32.
Schwarz, H.-G., Lang, C., Meier, S., 2007. Market power in the german
wholesale electricity market: What are the political options? Avail-
able at SSRN, http://ssrn.com/abstract=1028201 .
25
Sewalt, M., de Jong, C., 2003. Negative prices in electricity markets.
Commodities Now. Available at http://www.erasmusenergy.com/articles/91/1/Negative-prices-in-electrici%ty-markets/Page1.html .
Skytte, K., 1999. The regulating power market on the nordic power ex-
change nord pool: an econometric analysis. Energy Economics 21(4),
295–308.
Taylor, J. W., McSharry, P. E., 2007. Short-term load forecasting meth-
ods: An evaluation based on european data. IEEE Transactions on
Power Systems 22(4), 2213–2219.
Vehvilainen, I., Pyykkonen, T., 2005. Stochastic factor model for elec-
tricity spot price - the case of the nordic market. Energy Economics
27(2), 351–367.
26